Ask founders the question nobody puts in a vendor deck โ without generative AI, how many more employees would you need to produce last week's work at the same quality? โ and the average answer is a 55% headcount increase. The median is 17%. The most common answer is zero: 30% of founders who answered said they would need nobody extra (Colaiacovo & Koning, 2026).
Same question, same sample, three numbers that would justify three completely different budgets.
That is the first problem with how AI headcount savings get priced. The second is more useful: the gap between the founders reporting large savings and the ones reporting almost none is not explained by which tools they bought, or by how optimistic they are about what the models can do. It is explained by whether anyone restructured the work.
The Distribution Your Business Case Is Quoting
The study is Harvard Business School Working Paper 26-093 by Cassandra Colaiacovo and Rem Koning. They recruited 199 founders of North American technology ventures through Pitchbook โ 176 completed in full โ with firms averaging 16 employees (median 9), about 3.5 years old, over half venture-backed at an average of $4.86M raised. Ninety-six percent reported generative AI in use at the firm, formally or informally, with a median of five tools in play (Colaiacovo & Koning, 2026).
These are not reluctant adopters. This is close to the best case: small, flexible firms run by the person who owns both the workflow design and the hiring plan.
And the distribution of reported savings is heavily right-skewed. Mean 55%, median 17%, mode zero. The hours question behaves the same way: founders estimated needing 59% more work hours on average without AI, median 30%.
The operational consequence is specific. A business case built on the mean is a business case built on the tail. When a vendor or an internal champion says AI lets a team run at half the headcount, that figure is arithmetically true of the sample and descriptively false of the typical firm in it. The median firm in the most AI-saturated segment of the economy reports a 17% offset โ real, worth having, and nowhere near a layoff thesis.
If your agent plan depends on the mean, ask which half of the distribution you have evidence you are in.
Formal Integration Is Where the Headcount Savings Live
The authors asked founders to classify their own usage: was generative AI formally integrated into workflows, or used informally by employees as they saw fit? That single split reorders the results.
| Reported counterfactual without AI | Formal integrators | Informal users |
|---|---|---|
| Additional employees needed (mean) | 91.4% | 24% |
| Additional employees needed (median) | 40% | 10% |
| Additional work hours needed (mean) | 82.8% | 33% |
| Additional work hours needed (median) | 40% | 20% |
Nearly four times the headcount savings, on both the mean and the median (Colaiacovo & Koning, 2026). Formal integrators were also more likely to report that AI had changed whether and who they hire, made team coordination more efficient, supplied management capacity, and helped onboard new employees.
The authors are careful, and so should anyone citing them be: these estimates are associations, not causal effects. Founders who embed AI deliberately may be the same founders who would have found operating leverage anywhere. But the direction of the association is the actionable part, because the alternative explanation โ that formal integrators simply bought better tools โ is contradicted by what happens next in the paper.
What "Formal" Actually Meant
The distinction matters because the usual measurement fails at it. The authors note that asking which tools and foundation models a firm uses "can fail to distinguish between those using tools casually and those using them as part of regular work routines." So they let founders classify the thing that actually varies: whether generative AI sat inside the workflow or beside it.
That is the same measurement problem most mid-market AI dashboards have. Seat counts, weekly actives and prompt volumes all register the informal population at full strength. None of them can tell you whether a single process step was re-cut around the agent โ which is the only thing the formal/informal split is tracking, and the thing that moves the number by a factor of four.
Capability Beliefs Don't Explain Who Delegates
This is the result that moves the question out of procurement and into org design.
The researchers ran an embedded experiment on a function where AI deployment is already common: customer service. Founders were randomly shown either an optimistic or a pessimistic summary of real news and research about AI performance in that role, then asked again how willing they were to delegate a customer service inquiry to AI.
The treatment worked โ beliefs moved on average. Willingness to delegate did not converge. Task performance beliefs explained at most a third of the variation in willingness to delegate, and after the information treatment, revised beliefs explained only 37% (Colaiacovo & Koning, 2026). Notably, the deep integrators were not meaningfully more optimistic about AI's task performance than informal users. They were more willing to experiment with how humans and AI share the work.
Two-thirds of the variance in whether a capable agent gets real work sits somewhere other than the model.
That finding has company. Microsoft's 2026 Work Trend Index, surveying 20,000 workers across 10 countries, attributes more than twice as much of realized AI impact to organizational factors โ culture, manager support, talent practices โ as to individual ones, 67% against 32% (Microsoft Work Trend Index, 2026). Deloitte's survey of 3,235 leaders across 24 countries shows where the spend actually goes: 53% are educating the workforce to raise AI fluency, while only 30% are reimagining the organization around new AI patterns and 33% are redesigning career paths (Deloitte, 2026).
Everyone funds fluency. Almost nobody funds restructuring. The HBS data says restructuring is the variable that separates a 40% offset from a 10% one.
The Honest Counter
Three limits, stated plainly, because this is a survey and not a payroll audit.
These are beliefs, not measured output. Nobody reconciled founders' counterfactuals against headcount records or revenue per employee. The authors say so explicitly and call for work using realized productivity measures where balance-sheet data exists.
Beliefs about AI's headcount effect have a documented error rate. McKinsey's 2026 State of AI survey is the first edition able to grade its own panel: 32% of respondents a year earlier expected AI-related workforce reductions, and just 14% of AI-using organizations reported that AI contributed to an actual decline in workforce size over that period, with two-thirds reporting little or no AI-related change (McKinsey, 2026). Expectation overshot realization by roughly two to one. Read the HBS means through that lens and the skew gets more suspicious, not less.
The sample is 199 VC-backed tech startups averaging 16 people. A founder with four direct reports can redesign a workflow over a weekend. A 300-person operations function cannot, and its formal integration costs more and takes longer. The ordering of the findings travels to the mid-market; the magnitudes do not.
There is also a selection question running underneath all three. Formal integration is not free: it costs management attention, a redesigned process, and someone willing to own the output when an agent produces it. Firms that paid that cost are, by construction, firms that had the slack to pay it. For a 50โ500 FTE operation, the honest translation is that the four-fold gap is an upper bound on what redesign buys you, not a forecast of it.
None of this weakens the core claim. It narrows it: the separating variable is organizational, and the levels are self-reported.
What to Decide This Quarter
Four moves, none of which requires a new tool.
- Reprice the business case on the median, not the mean. If a plan assumes an offset above 40%, it assumes you are a formal integrator in the top decile of a right-skewed distribution. Write down which number you are underwriting and what evidence puts you there.
- Audit one workflow for formal integration, not for usage. Licence counts and weekly-active rates measure informal use โ exactly the population reporting a 10% median offset. The question is whether a process step has been re-cut around the agent, with an owner and a defined output.
- Separate the fluency budget from the redesign budget. They are not the same line item and they do not produce the same return. If training is funded and redesign is not, you are buying the informal-user outcome at the formal-integrator price.
- Give the agent a seat definition before you give it work. Formal integration means a named scope, a named accountable human, and an explicit rule for which actions need approval. That is the governance model behind Scovai's Digital Employees โ role-trained to a seat, confirm-first on anything externally visible, autonomy earned per kind of action rather than switched on โ useful as a reference for what "formally integrated" has to specify, whoever builds it.
The uncomfortable implication of this paper is that the agent capability question is mostly settled and the organizational one is mostly untouched. Ninety-six percent of these firms already had the tools. The ones reporting four times the savings had done something else.
Before you approve the next agent licence, find out who owns the workflow it is supposed to change. If the answer is nobody, you are buying the median.
